Everyone Will Have Their Own Agent
Every few weeks another company publishes a post about the agent it built for itself. Ramp has two. Shopify has one in Slack. Stripe has a fleet.
Stripe's engineers already have Claude Code, Cursor and Codex. They built their own anyway.
I've been doing the same at home with Hub. The more I read, the more I think everyone will have their own agent UI. As models get better, more people will build one from scratch, because a model can write it for them.
Companies build their own
Ramp's Inspect is a background coding agent wired into Sentry, Datadog, GitHub, Slack and Buildkite. You can reach it from Slack, a web app, a Chrome extension or the pull request. When Ramp wrote about it, Inspect wrote about 30% of merged pull requests in their frontend and backend repos, and nobody was forced to use it.
"Owning the tooling lets you build something significantly more powerful than an off-the-shelf tool will ever be. After all, it only has to work on your code."
Shopify's River works only in public Slack channels, so every session becomes a thread others can learn from. Stripe's Minions go from a Slack message to a pull request that passes CI, over 1,300 merged a week.
Same models, three different interfaces. Each one fits how that company already works.
Glass is the bridge
Ramp's second agent is the one I find most interesting. Seb Goddijn explains that Ramp hit 99% AI adoption and then most people got stuck. Terminals, npm installs and MCP configs were too much, and the few who pushed through had different setups with no way to share them.
So a small team built Glass. The company builds the base: every tool connected with one sign-in, a shared marketplace of more than 350 skills, memory built from your connected tools. Each person shapes the rest. You install the skills you need, schedule your own automations and arrange split panes and tabs, and the layout stays how you left it.
Glass is the shared-base path. A few people build the base, and everyone shapes their own on top. As Eric Glyman put it when Ramp opened Glass to other companies:
"No model ships with your company's brain."
Hub
I'm building from scratch. Hub runs 24/7 on a Mac Studio at home, and I open it on my phone over Tailscale.
The agents work on the Mac, and Hub mirrors their sessions. Every task is a card linked to the session doing the work, so a reply on the card goes straight to that session. The brief, my decisions, screenshots and results stay on the card even when the agent behind it changes. Chat is only for new requests.
What I wanted was one place showing what my agents are doing, what's waiting on me and what I already decided.
Building and shaping blur
Geoffrey Litt and his colleagues at Ink & Switch wrote the essay on this, Malleable Software:
"Malleable software does not imply everybody creating all of their own tools from scratch. That would be a waste of time, since many off-the-shelf tools are already pretty close to serving our needs, with only minor tweaks needed."
That held when you had to write the tool. A model can write it now. From scratch isn't the waste it used to be.
Maggie Appleton's barefoot developers are the people this is for: teachers with elaborate Notion setups, people with over-the-top personal dashboards. Technical enough to tinker, not interested in becoming programmers. A model writes the agent. They shape it around their use and their UI.
Why your own agent can move faster
When the only user is you, there's no release process. You ask for a change, it runs, and if it's wrong you roll it back. No PR, no review, no waiting on CI.
Models still get things wrong, but a mistake costs you a minute, not an outage. Products built for millions of people need guardrails. Something built for one person can skip them, and that's why it improves every day.
What to shape in yours
Whether you build or shape, these are the parts worth owning.
Where it lives. Start where you already are. Ramp and Stripe start in Slack, Karpathy's LLM Wiki in Obsidian, Litt's Stevens in Telegram.
The first view. Decide what you see when you open it. Mine is tasks waiting on me. Chats scroll away, and a task keeps its brief, status and result.
Rules. Write down what the agent can do alone and what waits for you. Anything that sends, pays, posts or deletes should wait for a tap.
Skills. When something works, save it as a markdown file your agent reads next time. That's what Ramp's 350 skills are, at company scale.
Memory you can read. Stevens is one SQLite table, Karpathy's wiki is markdown. If you can't read what your agent remembers, you can't fix it.
Senses, along with chat. Litt argues for AI HUDs, interfaces that help you see more. A status board can be worth more than another chat thread.
Start small. Anthropic's guide says to find "the simplest solution possible" and add complexity only when it clearly helps.
The hard parts
Security comes first. Simon Willison's lethal trifecta is private data, untrusted content and a way to send data out. An agent with your email has all three. Anyone can email it instructions.
Maintenance adds up. RevenueCat's post is titled Don't build your own harness. Moving to a standard coding loop removed over half of their system prompt. The harness is the plumbing. That part can live on a shared base, even if a model writes the rest.
Private agents have a ceiling. Shopify's point: if only you see your sessions, only you learn from them. Shared skills fix part of that.
Where I land
Everyone will have their own agent UI. As models get better, more people will build one from scratch, because a model can write it. Hub is that path for me. Glass is the shared-base path: sign-ins, mail, bank, memory, security and background runs sit on infrastructure you never see. The agent on top still feels built for you. Models are blurring the two.
Glass shows how the plumbing can be shared inside one company. I expect the from-scratch path to get easier for everyone else.